Plugin, Connector or MCP Server: What Is the Difference?

If you have tried to find a way to post on LinkedIn from an AI chat, you have probably searched at least three different phrases and got three different sets of results. That is not you being disorganised. One underlying thing genuinely has four names, and nobody has agreed on which one to use. This page is a decoder: what each word means, why the naming split, and which word to type where.

Are a plugin, a connector and an MCP server the same thing?

Nearly always, yes. All three describe the same arrangement: a piece of software that sits outside the AI assistant, and that the assistant is allowed to call when it needs to do something it cannot do on its own — look something up, send something, or publish something.

The difference is only which layer of the thing you are naming:

• Plugin names it from the user’s side — an add-on you plug into the app.
• Connector names it from the app’s side — the thing that connects the app to an outside account.
• MCP server names it from the builder’s side — the actual program, built to a shared standard called MCP.

Think of a phone charger. You can call it a charger, a power adapter, or a USB-C supply. Three names, one object, three different people describing it. None of them is wrong, and knowing all three saves you a lot of searching.

Why did the naming split in the first place?

The words arrived at different times, from different companies, and each one stuck with a different group of people.

“Plugin” came first and left a mark. It comes from an earlier generation of AI add-ons, and it is still the word most people reach for when they want an assistant to do something extra. It survives in how people search far more than in how the apps are labelled today.

“Connector” is what the product teams chose. When Claude and ChatGPT added the ability to attach outside tools to a chat, both landed on wording around connecting, because from your point of view you are connecting an account — your LinkedIn, your calendar, your files — not installing software. This is the word you will actually see in the menus.

“MCP server” is the engineering name. MCP stands for Model Context Protocol, an open standard published by Anthropic that describes how an AI app and an outside tool should talk to each other. The tool half of that conversation is called a server. So developers, documentation and code repositories all say “MCP server”, because that is what the thing is, regardless of which app is calling it.

“Integration” and “extension” are the inherited words. Business software has said “integration” for decades, and browsers have said “extension”. Both get borrowed here loosely, and both usually point at the same idea.

Which word does each app use?

This is the part worth bookmarking. Same object, different label depending on who is talking.

What is MCP, actually?

MCP — the Model Context Protocol — is an open standard for connecting AI assistants to outside tools. Open standard means the rules are published for anyone to read and build to, and no single company has to approve your tool before it can work.

Before a standard existed, every AI app had its own private way of attaching tools. If you built something for one app, you rebuilt it for the next. MCP replaced that with one agreed format, in roughly the way a standard plug socket replaced a different plug for every appliance.

The standard covers three things, in plain terms:

• Tools — the actions the assistant is allowed to take, each with a name and a described set of inputs. “Publish a post” is a tool.
• Resources — things the assistant is allowed to read.
• Prompts — ready-made starting instructions the tool can offer, so you do not have to invent the first message.

When you connect a server, the app asks it “what can you do?”, and the server answers with that list. That is why a connector can gain new abilities without you reinstalling anything — the list is fetched fresh each time, not baked into the app.

What does “server” mean here? I am not installing anything

This is the word that confuses people most, because “server” sounds like a machine humming in a cupboard. In MCP it simply means the side of the conversation that offers the tools. The app you are chatting in is the client. The tool is the server. That is the whole meaning.

There are two ways a server can reach you, and this distinction matters far more than the plugin-or-connector question:

• A local server runs on your own computer. You install it, usually with developer tools, edit a configuration file, and keep it updated yourself. It stops working when your laptop is off.
• A remote server runs on the web, and you connect to it by pasting a web address into the app. Nothing to install, nothing to update, and it keeps running when your laptop is closed — which is what makes scheduled posting possible at all.

Most of the LinkedIn tools you will find on code-sharing sites are local servers you run yourself. Kalovio is a remote one: you add https://kalovio.com/mcp and sign in. Same standard, very different amount of work.

Why can one server work in more than one app?

Because the standard is the contract. If a server speaks MCP correctly, any app that also speaks MCP can talk to it, without the server being rewritten for each one. That is the practical payoff of a standard, and it is why the same LinkedIn tool can appear in one app’s connector list and another app’s tools menu.

Two honest caveats, though, because “standard” does not mean “identical everywhere”:

• Apps support different parts of the standard. Tools are widely supported. The optional extras — things like a small interactive panel drawn inside the chat — are handled differently by different apps, or not at all. A server can work properly in two apps and still look different in each.
• Each app connects separately. Adding a server in one app does not carry it into another, even when both come from the same company. You add it and sign in again in each place.

Where do integration, extension, API and bot fit?

Four more words you will meet in the same search results:

• Integration — usually a fixed wiring between two products, built by one of the vendors. The closest cousin to a connector, but it normally does not involve an AI assistant deciding when to use it.
• Extension — software installed into an app and running on your machine, as described above. Different mechanism, so worth keeping separate.
• API — the application programming interface, meaning the door a service opens for other software. This sits one layer below every word above. A LinkedIn MCP server is, underneath, a program calling LinkedIn’s API on your behalf. The API is the door; the server is the thing that walks through it.
• Bot — a loose word usually implying software acting on its own with nobody watching. Worth avoiding here, because the tools we are describing act only when you ask them to, and LinkedIn treats unattended automation very differently.

Does the word change what you can do on LinkedIn?

Not at all. What you can do is decided by LinkedIn’s permissions, not by vocabulary, and those limits apply to every tool no matter what it calls itself:

• Publishing your own posts is available through LinkedIn’s free self-serve products, using the permission named w_member_social.
• Reading posts a tool did not create is not available. That permission, r_member_social, is closed to self-serve apps.
• Engagement analytics, comments and company-page posting sit behind LinkedIn’s partner-gated Community Management API. A tool offering those either holds partner status or is getting them some other way.
• Sign-in expires. Member tokens last around 60 days, and self-serve apps get no refresh token, so you reconnect from time to time.

So if two products describe themselves with different nouns but use the same self-serve permissions, they face exactly the same ceiling. The label on the tin tells you nothing about what is inside it.

Which word should I search for?

A short rule of thumb. If you want to know how to add it, search “connector”, because that is the word in the menus. If you want to know what exists, search “MCP server”, because that is what builders name their work. If you search “plugin”, you will land on pages like this one, written by people who worked out that the vocabulary is the actual problem.

And when you are comparing two tools, ignore the noun entirely. Ask instead: does it sign in through LinkedIn’s own consent screen, does it run remotely so scheduling still works when your laptop is shut, does it show you the exact text before publishing, and can it name the permission it uses. Those four answers tell you far more than any label.